Brickognize: Applying Photo-Realistic Image Synthesis for Lego Bricks Recognition with Limited Data.

During the last few years, supervised deep convolutional neural networks have become the state-of-the-art for image recognition tasks. Nevertheless, their performance is severely linked to the amount and quality of the training data. Acquiring and labeling data is a major challenge that limits their expansion to new applications, especially with limited data. Recognition of Lego bricks is a clear example of a real-world deep learning application that has been limited by the difficulties associated with data gathering and training. In this work, photo-realistic image synthesis and few-shot fine-tuning are proposed to overcome limited data in the context of Lego bricks recognition. Using synthetic images and a limited set of 20 real-world images from a controlled environment, the proposed system is evaluated on controlled and uncontrolled real-world testing datasets. Results show the good performance of the synthetically generated data and how limited data from a controlled domain can be successfully used for the few-shot fine-tuning of the synthetic training without a perceptible narrowing of its domain. Obtained results reach an AP50 value of 91.33% for uncontrolled scenarios and 98.7% for controlled ones.

Semantic Mapping for Autonomous Subsea Intervention.

This paper presents a method to build a semantic map to assist an underwater vehicle-manipulator system in performing intervention tasks autonomously in a submerged man-made pipe structure. The method is based on the integration of feature-based simultaneous localization and mapping (SLAM) and 3D object recognition using a database of a priori known objects. The robot uses Doppler velocity log (DVL), pressure, and attitude and heading reference system (AHRS) sensors for navigation and is equipped with a laser scanner providing non-coloured 3D point clouds of the inspected structure in real time. The object recognition module recognises the pipes and objects within the scan and passes them to the SLAM, which adds them to the map if not yet observed. Otherwise, it uses them to correct the map and the robot navigation if they were already mapped. The SLAM provides a consistent map and a drift-less navigation. Moreover, it provides a global identifier for every observed object instance and its pipe connectivity. This information is fed back to the object recognition module, where it is used to estimate the object classes using Bayesian techniques over the set of those object classes which are compatible in terms of pipe connectivity. This allows fusing of all the already available object observations to improve recognition. The outcome of the process is a semantic map made of pipes connected through valves, elbows and tees conforming to the real structure. Knowing the class and the position of objects will enable high-level manipulation commands in the near future.

H-SLAM: Rao-Blackwellized Particle Filter SLAM Using Hilbert Maps.

Occupancy Grid maps provide a probabilistic representation of space which is important for a variety of robotic applications like path planning and autonomous manipulation. In this paper, a SLAM (Simultaneous Localization and Mapping) framework capable of obtaining this representation online is presented. The H-SLAM (Hilbert Maps SLAM) is based on Hilbert Map representation and uses a Particle Filter to represent the robot state. Hilbert Maps offer a continuous probabilistic representation with a small memory footprint. We present a series of experimental results carried both in simulation and with real AUVs (Autonomous Underwater Vehicles). These results demonstrate that our approach is able to represent the environment more consistently while capable of running online.